shangeth/ljspeech-mimi-codes

LJSpeech — Mimi Codes

Pre-extracted Kyutai Mimi neural-codec tokens for the LJSpeech corpus — 13,100 English utterances from a single female speaker reading public-domain audiobook passages (~24 hours).

This dataset contains codes only, not audio. For waveforms, go to the original LJSpeech release; these codes are designed to be loaded alongside it for training Mimi-based speech models without paying the ~1 hour of GPU extraction cost.

Schema

One row per utterance:

Column	Type	Notes
id	string	e.g. LJ001-0001
text	string	normalized transcript, original mixed-case preserved
codes	int16[k=8][n_frames]	Mimi codebook indices @ 12.5 fps
n_frames	int32	= codes.shape[1]
k_codebooks	int32	= 8

Extraction details

• Codec: kyutai/mimi @ 24 kHz, 12.5 fps
• Codebooks: all 8 extracted. Slice codes[:k] if you want fewer (Mimi's codebooks are ordered by importance; the first few capture most of the signal).
• Codebook size: 2048 per codebook → values stored as int16
• Transcripts: the normalized column from metadata.csv (punctuation preserved, expanded numerics/abbreviations). Original mixed-case is kept — apply .lower() at load time if your model expects lowercase (e.g. to reproduce Wren-TTS training).

Usage

from datasets import load_dataset
import torch

ds = load_dataset("shangeth/ljspeech-mimi-codes", split="train")

ex    = ds[0]
codes = torch.tensor(ex["codes"], dtype=torch.long)   # [8, n_frames]
print(ex["id"], "→", ex["text"][:60])
print("codes:", codes.shape, "duration:", codes.shape[1] / 12.5, "s")

# Use only the first 3 codebooks (e.g. for a smaller model):
codes_3 = codes[:3]

Decode back to audio with the Mimi decoder:

from transformers import MimiModel
mimi = MimiModel.from_pretrained("kyutai/mimi").cuda().eval()
with torch.no_grad():
    wav = mimi.decode(codes.unsqueeze(0).cuda()).audio_values[0].cpu()
# wav is [1, T] @ 24 kHz

Splits

Split	Rows
train	~13,100

LJSpeech has no canonical train/val/test split — partition as your task requires.

License

The underlying LJSpeech corpus is in the public domain (CC0). The derived Mimi codes inherit this license. You can use, redistribute, and modify without attribution, though citing the original corpus is encouraged.

Links

• Dataset extraction code: github.com/shangeth/wren-datasets
• Wren research project: github.com/shangeth/wren
• TTS models trained on these codes: github.com/shangeth/wren-tts

Citation

@misc{wren2026,
  title  = {Wren: A Family of Small Open-Weight Models for Unified Speech-Text Modelling},
  author = {Shangeth Rajaa},
  year   = {2026},
  url    = {https://github.com/shangeth/wren}
}

@misc{ito2017lj,
  title  = {The LJ Speech Dataset},
  author = {Keith Ito and Linda Johnson},
  year   = {2017},
  url    = {https://keithito.com/LJ-Speech-Dataset/}
}

Related

Used to train the Wren series of speech-text multimodal models.